## Retrying Requests

Sometimes connections fail: either a pooled connection was stale and disconnected, or the webserver itself couldn’t be reached. OkHttp will retry the request with a different route if one is available.

## [Calls](https://square.github.io/okhttp/4.x/okhttp/okhttp3/-call/)

    try (Response response = client.newCall(request).execute()) {
      // Consume and discard the response body.
      response.body().source().readByteString();
    }

    System.out.println("REQUEST 2 (pooled connection)");
    try (Response response = client.newCall(request).execute()) {
      // Consume and discard the response body.
      response.body().source().readByteString();
    }
  }

 * follow-up requests. The client is unblocked and free to continue as soon as it has received the
 * entire response body. If and when the client makes a subsequent request using a pooled socket the
 * server may not have had time to close the socket. The socket will be closed at an indeterminate
 * point before or during the second request. It may be closed after client has started sending the

If there's a problem with the connection, OkHttp will select another route and try again. This allows OkHttp to recover when a subset of a server's addresses are unreachable. It's also useful when a pooled connection is stale or if the attempted TLS version is unsupported.

  }

  /**
   * Sets the delegate of `timeout` to [Timeout.NONE] and resets its underlying timeout
   * to the default configuration. Use this to avoid unexpected sharing of timeouts between pooled
   * connections.
   */
  private fun detachTimeout(timeout: ForwardingTimeout) {
    val oldDelegate = timeout.delegate
    timeout.setDelegate(Timeout.NONE)
    oldDelegate.clearDeadline()

## Version 4.4.1

_2020-03-08_

 *  Fix: Don't reuse a connection on redirect if certs match but DNS does not. For better
    locality and performance OkHttp attempts to use the same pooled connection across redirects and
    follow-ups. It independently shares connections when the IP addresses and certificates match,
    even if the host names do not. In 4.4.0 we introduced a regression where we shared a connection

	g := errgroup.WithNErrs(len(disks))
	for index := range disks {
		index := index
		g.Go(func() error {
			di := madmin.Disk{
				Endpoint:  endpoints[index].String(),
				PoolIndex: endpoints[index].PoolIdx,
				SetIndex:  endpoints[index].SetIdx,
				DiskIndex: endpoints[index].DiskIdx,
				Local:     endpoints[index].IsLocal,
			}
			if disks[index] == OfflineDisk {
				di.State = diskErrToDriveState(errDiskNotFound)

        when {
          client.fastFallback -> FastFallbackExchangeFinder(routePlanner, client.taskRunner)
          else -> SequentialExchangeFinder(routePlanner)
        }
    }
  }

  /** Finds a new or pooled connection to carry a forthcoming request and response. */
  internal fun initExchange(chain: RealInterceptorChain): Exchange {
    this.withLock {
      check(expectMoreExchanges) { "released" }

	if err != nil {
		t.Fatalf("NewEndpoint failed %v", endpoint)
	}

	if err = endpoint.UpdateIsLocal(); err != nil {
		t.Fatalf("UpdateIsLocal failed %v", err)
	}

	endpoint.PoolIdx = 0
	endpoint.SetIdx = 0
	endpoint.DiskIdx = 0

	poolEps := []PoolEndpoints{{
		Endpoints: Endpoints{endpoint},
	}}
	poolEps[0].SetCount = 1
	poolEps[0].DrivesPerSet = 1

    // Second call attempts the pooled connection, and it fails. Then it retries a new route which
    // succeeds.
    val response2 = client.newCall(request).execute()
    assertThat(response2.body.string()).isEqualTo("abc")
    assertThat(server.takeRequest().sequenceNumber).isEqualTo(1) // Pooled connection.
    assertThat(server.takeRequest().sequenceNumber).isEqualTo(0) // New connection.